Bayesian based Facial Expression Recognition Transformer Model in Uncertainty

Facial expression recognition (FER) has always been a major researching area. Accurate and robust FER system remains challenging due to the environment influence of human face, such as posture transformation, light illumination and occlusion. Besides, facial expression itself is extremely complex with emotion overlapping which determines that the expression dataset is inevitable with mislabeled or uncertain data. In this paper, we propose a new Transformer based architecture for the FER task combined with Bayesian theory. We also modify the feature extractor module and training strategy (namely Adapted-SCN) against the uncertainty from training data. Our novel architecture improves the performance by up to 90.86% accuracy on FERPlus and 74.69% accuracy on FER2013, respectively.